Traditional physical-based models have generally been used to model the resistive-switching behavior of resistive-switching memory (RSM). Recently, vacancy-based conduction-filament (CF) growth models have been used to model device characteristics of a wide range of RSM devices. However, few have focused on learning the other-device-parameter values (e.g., low-resistance state, high-resistance state, set voltage, and reset voltage) to compute the compliance-current (CC) value that controls the size of CF, which can influence the behavior of RSM devices. Additionally, traditional CF growth models are typically physical-based models, which can show accuracy limitations. Machine learning holds the promise of modeling vacancy-based CF growth by learning other-device-parameter values to compute the CC value with excellent accuracy via examples, bypassing the need to solve traditional physical-based equations. Here, we sidestep the accuracy issues by directly learning the relationship between other-device-parameter values to compute the CC values via a data-driven approach with high accuracy for test devices and various device types using machine learning. We perform the first modeling with machine-learned device parameters on aluminum-nitride-based RSM devices and are able to compute the CC values for nitrogen-vacancy-based CF growth using only a few RSM device parameters. This model may now allow the computation of accurate RSM device parameters for realistic device modeling.

Resistive-switching memory (RSM) is one of the most promising candidates for next-generation edge computing devices due to its excellent device performance. Currently, a number of experimental and modeling studies have been reported to understand the conduction behaviors. However, a complete physical picture that can describe the conduction behavior is still missing. Here, we present a conduction model that not only fully accounts for the rich conduction behaviors of RSM devices by harnessing a combination of electronic and thermal considerations via electron mobility and trap-depth and with excellent accuracy but also provides critical insight for continued design, optimization, and application. A physical model that is able to describe both the conduction and switching behaviors using only a single set of expressions is achieved. The proposed model reveals the role of temperature, mobility of electrons, and depth of traps, and allows accurate prediction of various set and reset processes obtained by an entirely new set of general current-limiting parameters.